Pangram's new detector sniffs out AI images and videos with heatmaps

Telling whether an image or video on your feed is real has officially become impossible for the human eye. Detection tools haven't been much better, often giving vague percentage scores that break the moment a file gets compressed.

AI detection startup Pangram just took its tools out of research preview with the launch of Pangram Image 1.0 and a lighter, speedier Pangram Image Flash 1.0. Both models are built to catch content from frontier AI generators across images and video.

Spotting edits with heatmaps and tackling noisy feeds

Instead of just spitting out an overall probability score, Image 1.0 can pinpoint the exact regions of an image that were synthetic. Its heatmap output shows manipulated spots even when an AI element is spliced into a genuine photograph.

To prove its chops, Pangram created 1,200 synthetic mirror images across 100 real photos using 12 frontier image generators, matching prompts to the real frames. The model hit 99.96% accuracy on that dataset, alongside benchmarks like X-AIGD, Mirage-Test, and AIGenImages2026.

More importantly for day-to-day use, the model is trained to withstand social media compression, noise, blur, and resolution drops. While the preview version refused anything below 512x512 pixels, Image 1.0 drops that floor down to 256x256, adding a lower-confidence disclaimer for tiny files rather than flat-out failing.

Frame-by-frame video forensics

Video detection is where things usually fall apart. Recent academic findings like DF26 show that humans perform no better than a coin toss when identifying AI talking-head clips, and conventional detectors lag behind.

That is why Pangram introduced Image Flash 1.0. On the DF26 benchmark, it notched a 98.7% accuracy rate, a massive 29-point jump over the closest competing detector. It also posted top results across benchmarks like AIGVDBench and Ivy-Fake.

To stress-test it against industry options, the team took 129 real social videos and generated synthetic mirrors across 14 frontier video generators, evaluating 1,529 videos in total. By feeding the generator the first real frame and a VLM description of the scene, they produced pairs with identical context to eliminate detector bias. Pangram now allows native video uploads on its site, analyzing footage frame by frame to show which seconds are synthetic.

Both models are accessible right away through an updated API and a Chrome extension that flags AI media as you browse feeds, rental listings, or shopping sites.

Why it matters

For anyone scrolling the web today, this is an immediately useful layer of defense. Between fabricated rental listings, fake product reviews, and synthetic talking heads pushing scams, relying on human gut checks is already a losing battle. Having a browser extension that can highlight suspicious patches gives users a practical sanity check before making decisions.

For platform operators and developers, the updated API offers a plug-and-play way to triage submissions without building forensic pipelines from scratch. However, the cat-and-mouse game is far from solved. Generation models move fast, and fraudsters will inevitably tune their outputs or add adversarial noise to bypass detectors once they become common knowledge.

We think Pangram's smartest design choice here is the heatmap and localized flagging rather than a blunt binary yes-or-no label. Instead of demanding absolute trust in an opaque confidence score, it shows its work and points your eyes to the anomaly. In an era where perfect detection is a pipe dream, equipping humans with interpretable visual clues is probably the most sensible path forward.